Page 211 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

#### bjt2

##### Senior member
sorry but can you explain this boot time calibration?
At boot time the CPU, since Carrizo, perform a measure of the TRUE Vcore erogated by the VRMs both without loading the CPU and with a fixed load. So the CPU calculates the DC offset and AC coefficient to compensate, later, the differences with the theoretical required Vcore. E.g.: the CPU orders to the VRMs to give 1V and the CPU measures the TRUE Vcore given by the VRMs both with ilde CPU and a fictious load (maybe one on more cores executing predefined power hungry code). With this, during normal operation, whenever the CPU decides that it needs X volt, it calculates the actual Vid that it must give to the VRMs to have exactly X volts, so you can lower the Vcore margins that you usually have and avoid malfunctions in case the VRM shift is over the tolerance.
EDIT: this also helps loose tolerance margins on the VRMs, meaning cheaper motherboards: if the VRM tolerance is higher, you can use low quality VRMs...
This calibration also compensate for VRM, PSU, CPU and MB ageing, that imply VRM voltage shifts during the years...

BTW also Fiji has boot time calibration.

#### KTE

##### Senior member
I answered it many times. The famous papers on ieeexplorer. I should save in the bookmarks the link to the post instead of writing it dozen of times...

AVFS construct periodically (probabily more than 100 times a second) a table of optimal frequency/voltage based on statistics collected on the 1300 replica circuits. Then calculates the Vmax applicable to the CPU based on the temperature and the CPU usage: there is a function called FIT that estimate the failure probability based on those two parameters and the Vmax is defined as the maximum voltage that let this FIT function be under a threshold of dependability.

Given power draw, temperature, this calculated Vmax and the frequency/voltage table calculated below, the CPU increase its frequency (and set the voltage accordingly to the table) if the power draw is under the limit, the temperature is under the limit and the required Vcore for this frequency is under the calculated Vmax. If the situation changes and the frequency is not sustainable anymore (one on more violations of those rules), the frequency will be lowered.

I suspect that TDP and temperature limits can be configurable and a Vmax offset (at your own risk) can be specified...

Let's see how many times further i must write these information...

EDIT: for the time skew: on Bristol Ridge the STAPM uses a filtered power limit, with an exponential mean, and acknowledge and higher boost state if the FILTERED power is under the limit. So this could also be the case for XFR... We will see...
Thanks, but you didn't answer 6/8 questions there.

I'll reply when I get time.

Or better, can you link the paper? I'll most likely have access.

Sent from HTC 10
(Opinions are own)

#### bjt2

##### Senior member
Thanks, but you didn't answer 6/8 questions there.

I'll reply when I get time.

Or better, can you link the paper? I'll most likely have access.

Sent from HTC 10
(Opinions are own)
For 7 and 8, i suppose that the temperature measure are spread all over the CPU and the maximum is taken. So even if you are drawing 20W because you are using only one core, you can't up the frequency too much, because that core will be the hot spot and will have too much temperature.
I think that there are not priorities on the conditions, meaning that the frequency is upped only if ALL contitions are met, namely the (filtered) power consumption MUST be under the configured TDP AND the (filtered?) temperature MUST be under the maximum temperature AND the needed Vcore for the frequency MUST be under the current calculated Vmax.

The papers are more than one:

First: Vdroop protection on Carrizo. http://ieeexplore.ieee.org/document/6757358/
Second: Bristol Ridge technologies (STAPM, AVFS, LDO, BTC etc) http://ieeexplore.ieee.org/document/7756316/
Third: Detailed explaination of Carrizo frequency/voltage table http://ieeexplore.ieee.org/document/7435018/ AND http://ieeexplore.ieee.org/document/7222485/

I'have not read all of them because i found and read the short paper for the ISSCC first for some of them, but these are the full papers...

#### KompuKare

##### Senior member
reminds me of fury x fiasco. AMD: oh this cooling is great for overclocking! turned out you couldnt OC shit.
But was that due to the cooling?
I know that AMD got a lot slack for that statement, but in retrospect I read it as them saying their reference board was way over-engineered. However, the forgot to say that the actual chip only had about 5% margin left.
"Board willing, chip unable" whereas the competitor's reference boards are often "chip able, board unable". Previous AMD reference cards were more "chip able, board able, cooling unable"...

#### bjt2

##### Senior member
For 7 and 8, i suppose that the temperature measure are spread all over the CPU and the maximum is taken. So even if you are drawing 20W because you are using only one core, you can't up the frequency too much, because that core will be the hot spot and will have too much temperature.
I think that there are not priorities on the conditions, meaning that the frequency is upped only if ALL contitions are met, namely the (filtered) power consumption MUST be under the configured TDP AND the (filtered?) temperature MUST be under the maximum temperature AND the needed Vcore for the frequency MUST be under the current calculated Vmax.

The papers are more than one:

First: Vdroop protection on Carrizo. http://ieeexplore.ieee.org/document/6757358/
Second: Bristol Ridge technologies (STAPM, AVFS, LDO, BTC etc) http://ieeexplore.ieee.org/document/7756316/
Third: Detailed explaination of Carrizo frequency/voltage table http://ieeexplore.ieee.org/document/7435018/ AND http://ieeexplore.ieee.org/document/7222485/

I'have not read all of them because i found and read the short paper for the ISSCC first for some of them, but these are the full papers...

I was reading the detailed papers that i have not read and i found new details:
- The statistics is used to construct a voltage frequency temperature table (10 bit per entry, 10x8 entries, encoding the optimum voltage), so that the statistics must not be performed too often, because the table includes more operating points. The statistics calculate gaussians for various type of circuits and include guard bands. The circuit replica is checked for the results delay and the "near miss" are counted. Near miss are results that are arrived just in time and would be too late if only the Vcore is slightly lower.
- There is one table for each core so the Vcore is calibrated for each and so die inhomogeneity is taken into account (it is possible to have different Vcore for each core, see DLDO specification)

#### OrangeKhrush

##### Senior member
reminds me of fury x fiasco. AMD: oh this cooling is great for overclocking! turned out you couldnt OC shit.
I cannot say what the max overclock will be, and in all likelihood like the old Phenom II's looks like the OC range may not really go beyond 4.5 without high quality cooling.

Likely culprits:

1) 14nm FinFET LP yields, this is just process maturation, AMD and GLOFoundries are new to this process and yields are likely to improve on later steppings.

2) Firmware, again the process may take teething time, as long as there is clear attempt to remedy the instabilities that is acceptable

Either way Zen is very promising despite its few blemishes.

#### guskline

##### Diamond Member
My 5960x below is SOLID OCer at 4.4GHZ running Asus RealBench. I know there are other posters with higher OCs and perhaps my chip is mediocre; however I found this to be solid and stable.

2 years have passed and now a Ryzen 1800x is in my sights. I intend to OC it with a EK Supremacy EVO block and 2 360 slim rads which will fit in my Fractal Define S case. I'm anxious to see what clock speeds I get and if custom water cooling helps.

I agree that cranking these chips up produces more heat; 8 cores 16 threads will do that.

#### JoeRambo

##### Golden Member
Isn't that a horrible waste of money doing such sidegrade to Zen? You are moving from 8 core @4.4 to 8 core @ some unknown clocks, but you will hardly gain any measurable performance delta to justify it, unless Zen is running @5Ghz or so?

#### coercitiv

##### Diamond Member
Isn't that a horrible waste of money doing such sidegrade to Zen? You are moving from 8 core @4.4 to 8 core @ some unknown clocks, but you will hardly gain any measurable performance delta to justify it, unless Zen is running @5Ghz or so?
Take a look at his sig: it's a hobby, not a need&value oriented purchase.

#### JoeRambo

##### Golden Member
Hobby with very good selection of items. One that would better be served by getting SM961 or 960 PRO? Or 1080 TI when NV releases it?

#### Crumpet

##### Senior member
I'll be at least partially upgrading sideways from an i5 6600, as the difference in Single Thread is likely not going to be much even with Ryzen at higher clocks, though i'd like to be wrong.

Everyone has their own reasons for getting a Ryzen chip.

Mine just happens to be that I want to go back to supporting the underdog. I tried out Intel and Nvidia as it is always recommended as the superior product. Well, now that I have I have zero qualms with going back. I'd even take my fx8350 back from the wife if she'd let me, which she won't.
To me, as it stands today, AMD are practising better company standards than the two competitors named above, and I also feel that they are doing more to future the development of computing as a whole.

These are my reasons and my reasons alone.

#### flash-gordon

##### Member
My 5960x below is SOLID OCer at 4.4GHZ running Asus RealBench. I know there are other posters with higher OCs and perhaps my chip is mediocre; however I found this to be solid and stable.

2 years have passed and now a Ryzen 1800x is in my sights. I intend to OC it with a EK Supremacy EVO block and 2 360 slim rads which will fit in my Fractal Define S case. I'm anxious to see what clock speeds I get and if custom water cooling helps.

I agree that cranking these chips up produces more heat; 8 cores 16 threads will do that.
If you need an excuse, say that you were needing Integer performance...

#### guskline

##### Diamond Member
I don't need an excuse. It is a hobby. I have (had) 3 rigs. 5960x;6700k and 4790k. The 4790k components are now for sale while I prep the Fractal Define S case for the Ryzen 1800x/mb/ram combo.

I had 2 RX480s in CF running with the 6700k. They will now run with Ryzen for an all AMD cpu/gpu combo.

The 6700k was kept because it is a newer mb Asus M8H and I could use the custom water cooled GTX980TISC (EK block with it). The 6700k is using an EK Supremacy EVO block. I will also use that same block for the Ryzen 1800x (I already have the AM4 adapter).

I will not water cool the RX 480s in CF because they are reference and watercooling them isn't worth the cost of the blocks vs the cost of the gpus. I'll see whether Vega is worth water cooling later this year.

I intend to keep my 5960x and when the Ryzen 1800x rig is completed it will be fun to compared benchmarks.

As far as "horrible waste of money", it's hard to argue against someone feeling that way. I'm not a golfer and could argue that buying new clubs, shoes etc all the time is a "horrible wate of money". However, I will not do that since I have learned that if something is your hobby (i.e. golf to a golfer and building newer computers to me) and you love it, wasting money is probably not in the equation.

FORE!!!!!!!!!

Last edited:

#### Vaporizer

##### Member
I'll be at least partially upgrading sideways from an i5 6600, as the difference in Single Thread is likely not going to be much even with Ryzen at higher clocks, though i'd like to be wrong.

Everyone has their own reasons for getting a Ryzen chip.

Mine just happens to be that I want to go back to supporting the underdog. I tried out Intel and Nvidia as it is always recommended as the superior product. Well, now that I have I have zero qualms with going back. I'd even take my fx8350 back from the wife if she'd let me, which she won't.
To me, as it stands today, AMD are practising better company standards than the two competitors named above, and I also feel that they are doing more to future the development of computing as a whole.

These are my reasons and my reasons alone.
Its not your reasons alone. I share the same reasons

Crumpet

#### guskline

##### Diamond Member
ShintaiDK, is that post concerning the 4c/8t Ryzen or the 8c/16t Ryzen? The threads I read seem to imply the 4c/8t Ryzen.

No argument that right now the KabyLake/Skylake 7700k/6700k, especially OC'd seemed to be the fastest gaming cpus.

Heck my 6700k at 6700k is slightly faster than my 5960x clock for clock but BOY OH BOY when you want or need 8c/16 t the 5960x digs in.

It will be interesting to see how close the Ryzen 1800x is to my 5960x in overall performance. If the quoted prices are correct (@\$500) it will be 1/2 the price. I realize the Ryzen is only dual channel so there will likely be some loss there but it will be interesting to see how it stacks up.

Last edited:

#### ShintaiDK

##### Lifer
ShintaiDK, is that post concerning the 4c/8t Ryzen or the 8c/16t Ryzen? The threads I read seem to imply the 4c/8t Ryzen.
If it was for 4/8, why would it then not be slower in highly MT?

#### OrangeKhrush

##### Senior member
3.2/3.4 is strangely low, will definately have to push clockspeed up on the next arch, that said while it won't beat Kabylake which is pretty much known anyways, it will be good enough to compete with Haswell parts.

#### Arachnotronic

##### Lifer
it will be good enough to compete with Haswell parts.
Haswell is nearly four years old.

#### lolfail9001

##### Golden Member
Haswell is nearly four years old.
To be honest, Intel did not go that far from these 3-4 year old parts either.

I don't like that resemblance of hard wall arising.

cytg111

#### Glo.

##### Diamond Member
Haswell is nearly four years old.
How much faster is KabyLake at the same clock, than Haswell, and how much Haswell CPUs with 8 threads cost in 2017, compared to rumors about Ryzen? And which architecture is more efficient?

#### OrangeKhrush

##### Senior member
The HTPC is a i5 4460 with 3.2/3.4 clocks, admittedly it is not great for intense gaming but it is far from what I would call bad. If I could get a 4/8 CPU with the same level of performance for a price less than a 7350K then it is a good deal. Still the clocks though need to raise, maybe with silicon maturity and steppings we will like the FX4100 was followed by the 4150 and 4300 maybe a 1350X or something will come out with higher frequency.

The only disappointing aspect is the clocks right now.

#### jpiniero

##### Lifer
On 4 core gaming performance. Not 8 core.
Unless the 6/8 core models clock much higher the gaming performance difference in most games won't be much between the 4C8T Ryzen and the higher ones.

Status
Not open for further replies.